So, when Claude Code came out, I decided to lay off writing coding agents for a while. I figured, there were already a dozen pretty good coding agents, and at some point, someone would figure out everything I've figured out, and finally do one right.
Figured, I'll wait six months... see what happens.
The state of the technology hasn't really moved.
Everyone's making more or less the same mistakes, the state of the human in the loop coding agent has not improved, which is weird... because it did for a minute. We had a lot of real innovation there for a second.
None of them have quite gotten memory right. MCP isn't the answer for memory systems, and the closest thing to correct is ticket system based memory schemes, which seem like they should work... but still fall short of the kind of thing I've been envisioning.
Also, none of them have sufficient repo safety built in. I'm going to make it impossible for an agent to melt your project or burn your code.
Thinking, it's time to finish one of my human in the loop coding agent projects. I'll post notes as I go. Going to shoot for having it done this weekend.